DeepSeek V4 Flash

PROVISIONAL CALIBRATION per /NM310-A1 (2026-09-07) — 0 BAND / 2 BAND-COMPOSITE / 17 INHERITED-PROVISIONAL / 1 PROXY; re-calibrate when per-dimension evals land. Released 2026-04-24, MIT license. Sparse MoE: 284B total / 13B active params (~160GB — self-hostable on high-memory Apple Silicon or multi-GPU workstation once quantized). Context 1M (1,048,576), 384K max output — unusual long-output headroom. Reasoning + non-reasoning modes. List pricing $0.14/$0.28 per M (provider floor ~$0.09/$0.18; cache hits ~$0.0028/M — near-free input on repeated-context agent loops). AA measurements: Intelligence Index 46.5 / Coding Index 38.7 (high-reasoning, 2026-05-12 snapshot). Highlighted by AI Explained (2026-07-10) as a SimpleBench performance-per-dollar Pareto standout: ~50% at extremely low cost. Ships with DSpark speculative-decoding variant on Hugging Face. Parent Model left empty: no DeepSeek V4 family/Pro record in DB yet (DB has V3.2, R1) — create DeepSeek V4 Pro record if the family gets tracked further. PROVENANCE (-A1 lanes M1/M1b, per-dim): Reasoning + Bulk Apperception [BAND-COMPOSITE:DeepSeek V4 Flash — AA Intelligence Index (2026-05) 46.5]; Cost Efficiency [PROXY:AA $0.175/M blended]; all remaining 17 dimensions [INHERITED-PROVISIONAL:deepseek-v3-2]. Code Generation is inherited (14): AA Coding Index 38.7 is on file but the adjudicators rejected banding it (its Benchmarks Calibration Stance is Display-Only per, and the candidate value 7 was a 7-point downgrade vs the parent), so A1 clause 3 fallback applies. Parent Model wired to DeepSeek V3.2 as inheritance anchor pending a V4 Pro record.